A method of processing an image

Abstract

A method for processing a seed image, the method comprising: inputting a seed image comprising at least a portion of a seed to a trained neural network to generate a value relating to coating coverage on the seed; the trained neural network having been trained to generate values relating to coating coverage on seeds using a plurality of training images, each training image comprising at least a portion of a training seed, and wherein each training image is labelled with a value relating to coating coverage on the training seed.

Description

Technical Field

[0001] The present invention relates to a method for processing seed images, and particularly, but not exclusively, to a method for processing seed images including a coating. Background Technology

[0002] Seeds used in agriculture are often coated. Coating can be provided for various reasons. Generally, seed coating protects the seeds from damage during handling, prevents dust accumulation, and enhances their appearance. This coating also provides protection against pests and diseases, as well as smoothing the seed surface to make planting easier. To control seed germination or germination rate, phytonutrients or other growth stimulants can be incorporated into the seed coating. Plant protectants, such as pesticides (e.g., fungicides and insecticides), can be added to further protect the seeds from diseases and / or pests.

[0003] Seeds are frequently subjected to mechanical stress, especially during processing. Therefore, the seed coating is also subjected to mechanical stress. Advantageously, the seed coating resists abrasion caused by mechanical stress, thus maintaining adequate coating and enabling the coating to perform its intended function, making the seed visible in the soil and / or ensuring the coating remains visible as a means of identifying seed type and its specific coating.

[0004] Abrasion resistance is the ability of a seed coating to resist mechanical stress and is a measure of the quality of the film coating on seeds. Visual inspection of seeds can be used to determine the abrasion resistance of the coating, particularly by assessing the degree of coating damage after the seeds have been subjected to abrasion. However, these methods can be subjective because many random factors cannot be fully controlled (e.g., personal preference, lighting, or visual reference), leading to significant variations even in identical samples. There is a desire to develop more accurate and consistent methods for assessing coating abrasion resistance.

[0005] Invention Statement

[0006] According to one aspect of the present invention, a method for processing a seed image is provided, the method comprising: inputting a seed image including at least a portion of a seed into a trained neural network to generate a value related to the coating coverage on the seed; the trained neural network has been trained to generate the value related to the coating coverage on the seed using a plurality of training images, each training image including at least a portion of a training seed, and wherein each training image is labeled with a value related to the coating coverage on the training seed.

[0007] Values ​​associated with coating coverage on seeds can indicate the abrasion resistance of the coating. Values ​​associated with coating coverage on seeds / training seeds can indicate the percentage of seed surface covered by the coating. Values ​​associated with coating coverage on seeds can indicate the uniformity of coating coverage on the seeds. Values ​​associated with coating coverage on seeds can include representations of the uniformity of coating application to the seed surface and / or the degree to which the seed surface is covered by the coating. The amount of surface covered by the coating can indicate the abrasion resistance of the coating. For example, if the coated seed has been exposed to mechanical stress, the amount of coating remaining can indicate the abrasion resistance of the coating, or the thickness of the coating can indicate the abrasion resistance of the coating.

[0008] The method may further include a step of training a neural network. A first training image containing a training seed with a first color coating and a second training image containing a training seed with a second color coating can be used to train, or a trained neural network can already be trained. The input step may include inputting a seed image containing a seed with a first color coating.

[0009] A seed image can be generated by extracting the seed regions from an image containing multiple seeds. Alternatively, an algorithm that detects circles in an image can be used to extract the seed regions, detecting the areas within the seeds and then extracting those areas.

[0010] Training images can be generated by extracting regions of training seeds from an image containing multiple training seeds. Algorithms that detect circles in an image can be used to extract regions of training seeds, detecting regions within the training seeds and then extracting those regions.

[0011] Multiple seed images can be generated from the same image or taken from the same sample (e.g., the same seed or the same sample containing multiple seeds). A sample can be a collection of multiple seeds that have been coated with substantially the same amount (and type) of coating and exposed to substantially the same amount of mechanical stress. These images can be fed into a trained neural network to generate multiple values ​​related to the coating coverage of the corresponding seed image. The average of these values ​​can be taken.

[0012] A neural network can be at least one of the following: a convolutional neural network; and a deep neural network. A neural network can include at least one of the following (each): an input layer, a 2D convolutional layer, a batch normalization layer, a ReLU layer, a max pooling layer, a fully connected layer, and a regression layer.

[0013] According to one aspect of the present invention, a computer program is provided that, when executed by a computing system including processor hardware and memory hardware, causes the processor hardware to perform any of the methods described herein.

[0014] According to one aspect of the present invention, an apparatus is provided comprising processor hardware and memory hardware, the memory hardware storing processing instructions that, when executed by the processor hardware, cause the processor hardware to perform any of the methods described herein.

[0015] To avoid unnecessary repetition and verbosity in the specification, certain features are described only for one or a few aspects or embodiments of the invention. However, it should be understood that, where technically possible, the features described with respect to any aspect or embodiment of the invention can also be used for any other aspect or embodiment of the invention. Attached Figure Description

[0016] To better understand the invention and to more clearly illustrate how to implement it, reference will now be made to the accompanying drawings by way of example, wherein:

[0017] Figure 1a Here is a flowchart of a method based on an example;

[0018] Figure 1b Here is a flowchart of a method based on an example;

[0019] Figure 2 An illustration of an image involved in an example image processing method;

[0020] Figure 3 This is an illustration of the seed part of a neural network, based on an example, using coverage values ​​(labels) attributed to them;

[0021] Figure 4 Image illustrations used for neural network validation;

[0022] Figure 5 An illustration of the seed portion used to validate the neural network, which has coverage values ​​and predicted values ​​attributed to them by the neural network; and

[0023] Figure 6 A graph illustrating the correlation between the coverage values ​​attributed to the network and the predicted values ​​generated by the neural network;

[0024] Figure 7 for Figure 2 The color version;

[0025] Figure 8 for Figure 3 The color version;

[0026] Figure 9 for Figure 4 A color version; and

[0027] Figure 10 for Figure 5The color version. Invention Details

[0029] Seed coatings are usually colored to indicate that the seeds have been treated. The specific color of the coating can indicate its composition. Alternatively, the color can be chosen to make the seeds more visible in the soil. Blue coatings are less noticeable in soil, while red coatings are more prominent. Seeds may also be coated with a color to reduce their attractiveness to animals. Crops such as corn or sunflowers are often coated with red or blue.

[0030] Colored seed coatings can be used to determine the coating's abrasion resistance (or its resistance to detachment or displacement from the seed under mechanical stress) because the coverage and / or thickness of the coating on the seed can be determined by comparing the color of the colored coating with that of the untreated seed. Therefore, seeds can be visually inspected and assigned a value to represent the coverage of the seed coating, thus indicating the coating's abrasion resistance and / or the percentage of the seed surface covered by the coating. Values ​​associated with seed coverage can indicate the uniformity of the coating application on the seed surface and / or the extent to which the coating covers the seed surface. The amount of surface coating and / or the thickness of the coating can represent the coating's abrasion resistance.

[0031] The abrasion resistance of seed coatings can be tested by applying mechanical pressure to the seed. For example, abrasion of the seed coating can be achieved by placing at least one seed in a wheel (roller, turntable) and rotating the wheel for a certain amount of time and / or revolutions. The wheel can rotate about a horizontal axis in a vertical plane. Alternatively, the wheel can rotate about a vertical axis in a horizontal plane. Any suitable device for applying mechanical pressure to the seed can be used. For example, an abrasion testing machine can be used. Examples of such machines include PharmaTest's PTF 200, PTF 20E, or PTF 20ER models (commonly used to test the brittleness or abrasion resistance of pharmaceutical tablets, but equally applicable to seeds). Abrasion can then be visually assessed by evaluating the remaining amount of colored coating, with values ​​within a range attributed to the coating. The coating may have been completely removed from the seed surface, or the coating thickness may have decreased (which may result in a different color or hue than a full-thickness coating). Seed abrasion (or exposure of the seed to mechanical pressure) not only results in material loss but also in a reduction in layer thickness or partial coating detachment. It can also cause the coating to shift to other parts of the seed. This can result in some areas of the coating becoming thinner or falling off completely, while in other areas the coating thickness may increase. This can lead to more uneven coverage that cannot be measured using methods such as gravimetric analysis, but can be measured by evaluating the coating thickness. Therefore, the abrasion resistance of the coating can be determined by evaluating the thickness of the coating on a seed exposed to mechanical stress. Lower values ​​in this range may indicate no abrasion (high coating coverage), while higher values ​​may indicate that most of the coating has fallen off. For example, a number between 0 (no abrasion / generally complete coating coverage) and 5 (almost no coating remaining) can be assigned to a seed or coating.

[0032] However, due to random factors that cannot be properly controlled (such as personal preference, lighting, or visual reference), visual inspection of seeds can be relatively subjective, and even in the same sample, it can lead to significant differences in the determined abrasion resistance.

[0033] In one example, a method for processing seed images is provided, the method comprising inputting a seed image including at least a portion of a seed into a trained neural network to generate a value associated with (or related to) the coating coverage on the seed. The trained neural network is a neural network trained to generate values ​​associated with the coating coverage on the seed using multiple training images, each training image including at least a portion of a training seed, and wherein each training image is labeled with a value associated with the coating coverage on the training seed. Figure 1a This method is outlined in the paper, which describes the steps of inputting a seed image, including at least a portion of the seed, into a trained neural network to generate a value related to the coating coverage on the seed.

[0034] This method can include training a neural network. Figure 1b This method is outlined in [the document]. Therefore, the method may include training a neural network using multiple training images to generate a trained neural network that generates values ​​related to the coating coverage on a seed, each training image including at least a portion of a training seed, and wherein each training image is labeled with a value related to the coating coverage on the training seed (S100). Seed images including at least a portion of the seed can then be input into the trained neural network to generate values ​​related to the coating coverage on the seed (S102).

[0035] The determined coverage can represent the percentage of the seed surface covered by the coating (compared to the percentage of the seed surface without coating or substantially without coating). The determined coverage can represent the uniformity of the coating application on the seed surface during the coating process. The determined coverage can indicate the uniformity (thickness) of the coating (e.g., after exposure to mechanical stress). The determined coverage can represent the abrasion resistance of the coating. As mentioned above, abrasion resistance is the ability of the coating to resist mechanical stress. The abrasion resistance of the coating on the seed can be tested by exposing the seed to mechanical stress and determining how much coating detaches (the amount of detached coating can be the amount of coating that detaches completely or substantially completely, and / or may include the amount of reduction (or increase) in coating thickness). Therefore, determining the coverage of the seed coating (e.g., coverage percentage, reduction (or increase) in thickness, etc.) after the seed has been exposed to mechanical stress can represent the abrasion resistance (wear resistance) of the coating. For example, values ​​associated with the coating coverage on the seed may be related to the seed's abrasion resistance.

[0036] This method can utilize images of the abraded surface of the seeds. For example, the seeds of a sample can be placed in a wheel that rotates for a certain period of time or number of revolutions, thereby subjecting the seeds to mechanical pressure. An image of the seeds can then be captured using a camera, scanner, or any other image extraction means.

[0037] For the images generated for training (training images), for the same type of seeds with the same type of coating, different numbers of times or rotations of the wheel can be used to achieve multiple seed coverage rates. Images of seeds that have been processed in this way can then be taken, resulting in images of seeds with different coating coverage rates. Images can also be taken of the same type of seeds that are coated but not exposed to mechanical stress, as well as unprocessed (uncoated) seeds. The images of seeds used in training can be images showing seeds with different coverage levels and can include coated seeds that have not been exposed to mechanical stress, coated seeds that have been exposed to mechanical stress, and / or uncoated raw seeds. Images can also be images of different qualities, resolutions, and / or light intensities. These images can be used as training images (e.g., training data) for training a neural network. Multiple training images can be reserved for validation of the method (validation images), for example, to check that the trained neural network is assigning the correct coverage / abrasion resistance values ​​to the images.

[0038] Furthermore, these images can be images of coated seeds, whose abrasion resistance will be determined. These images (seed images) can be input into a trained neural network. To determine the abrasion resistance of the coating, it may be desirable to establish a set time or number of revolutions for rotating coated seed samples in a wheel, so that different seed samples with different coatings (of the same type of seed) are subjected to substantially the same mechanical pressure, thus making the results of different types of coatings (on the same type of seed) comparable.

[0039] The captured seed images can be initial images comprising multiple seeds. These initial images can be used as input to a neural network or as input to a trained neural network. Alternatively, data about a single seed can be extracted from the initial images. For example, an image of a single seed or a portion of a single seed can be extracted from the initial images. This is further advantageous because the initial images can include multiple single seeds, thus increasing the amount of training data that can be input into the neural network by extracting single seeds. Furthermore, to determine coating coverage, partial images of a single seed can be grouped based on samples from which the single seed was produced (e.g., seeds coated with the same type of coating, from the same batch of coated seeds), and an average value representing the coverage can be taken. For example, images of a single seed portion taken from the same initial image or the same sample can be processed by a neural network to provide abrasion resistance values ​​individually, which are then averaged to generate an average value associated with the coating coverage of each sample. This can provide a more accurate abrasion resistance value for a particular coating.

[0040] Computer-based image processing methods can be used to extract data about seeds from images. For example, seeds (such as corn and tomatoes) often have a roughly circular or elliptical shape, so an algorithm based on the Circular Hough Transform (CHT) for finding circles in an image can be applied to identify the radius and center coordinates of the seed in the initial image. The circumference of the circle can be tangent to the edge of the seed. Therefore, the circle can be within the seed's perimeter. Alternatively, the circle can contain overlapping regions of the seed, where the circle overlaps with at least a portion of the seed. Note that this method can be applied to any type of seed to identify regions within the seed. These identified regions can then be extracted. This method is advantageous because it is stable in the presence of noise, occlusion, and varying seed illumination. This extraction method can also be applied to seed images input to a trained neural network to determine the coverage of the seed coat. The results of this method are as follows... Figure 2 As shown.

[0041] Figure 2 An image of corn 210 is shown, which has been coated (in a red coating) and exposed to the mechanical pressure described above. In the example described herein, the seed used is corn; however, it should be understood that this method will be applicable to images of any type of seed and / or coating color. The method described above for extracting regions of individual seeds is applied to the corn image 210, producing a processed image 212, in which regions detected as belonging to individual seeds are represented by circles. A portion of each individual seed (or a portion of a group of individual seeds) is then extracted. Figure 2 An example of a single extracted seed region 214 is shown. Specifically, the circular region represented in the processed image 212 has been extracted and is a portion of the image of a single seed. Any number of extracted regions can be used to train a neural network or as input to a trained neural network.

[0042] By processing the initial images in this way, multiple individual seed images can be generated to serve as training images. In one example, by processing 500 initial images in this way, approximately 30,000 individual seed images can be generated and used for training and validation (in such a set, 10% of the images can be reserved for validating the method).

[0043] Figure 3 Several such images are shown in the document. Figure 3The image shows partial regions of 25 individual seeds extracted from the initial image. These images are seed regions with different colored coatings, coverage (thickness and percentage), lighting, etc. Specifically, these images are images of blue or red coatings on yellow corn. As shown in the figure, each image is associated with a coverage value evaluated by visual inspection of the seed (the image number appears next to the value labeled "Measured," where the measured value is the coverage value evaluated by visual inspection of the seed). This value can be provided as a general value for seed coverage in an initial image with multiple seeds and can be applied to each extracted partial seed region extracted from the initial image (the same value attributed to the entire image can be applied to each seed region). This value can be a range of values. For example, values ​​between 0 (fully coated) and 5 (no coating) can be assigned to each training image as an indication of coating coverage on the seed. Since they are black and white copies, they are not visible in these images, but these (used in the method) images are color images, so areas where the coating has fallen off may appear as the color of the seed below. In this specific example, the color of the seed (corn) below is yellow, and the coating is blue or red. Therefore, the ratio of yellow to red or blue in an image may be related to the coverage of the seed coating. Thus, the ratio of an image including the color corresponding to the color of an uncoated seed to an image including the color corresponding to the coating applied to the seed can represent the coverage of the seed coating (e.g., how seeds are visually evaluated, and this analysis can be used to determine values ​​associated with training images). It should be noted that partial shedding of the coating, such as a decrease (or increase) in coating thickness, can result in a color between the coating color and the actual color of the seed, or it may be a different hue of the coating color, and such intermediate colors or different hues associated with partial wear of the coating can also be considered when generating values ​​associated with coating coverage. Therefore, values ​​associated with seed coating coverage can represent how much of the seed coating has shed (or how much remains), ranging from values ​​indicating no coverage to values ​​indicating complete coverage. Complete coverage can mean that the seed is essentially completely covered by a coating of a certain thickness (within the margin of error). For example, in the range of 0-5, a seed with 50% coverage can be represented by a value of 2.5. Values ​​associated with coating coverage can also be abrasion resistance scores. For example, an abrasion resistance score of 2.5 can indicate that the seed is covered by approximately 50% coating. It should be understood that grayscale images can also be used in the methods described herein, where the relative darkness of a region in the image can represent the seed coverage.

[0044] In one example, seeds with different colored coatings were used. Specifically, seeds with blue and red coatings were used, although it's understandable that other colored coatings or only one color coating could be used to train the neural network. Alternatively, a neural network could be trained using multiple different colored seed coatings. Using seeds with different colored coatings produced a surprising effect: when the neural network was trained with seeds of the first and second colors, rather than only with seeds of the first color, the neural network was able to identify the coverage of the first-colored seeds more accurately.

[0045] When discussing different colors, it can be assumed that each color is defined by a corresponding wavelength interval in the visible spectrum of that color. For example, the color "red" can be defined as having a wavelength interval of approximately 700-635 nm, "orange" as approximately 635-590 nm, "yellow" as approximately 590-560 nm, "green" as approximately 560-520 nm, "cyan" as approximately 520-490 nm, "blue" as approximately 490-450 nm, and "violet" as approximately 450-400 nm. When referring to "different colors," the color can be a color attributed to a different wavelength interval. Alternatively, "different colors" or "different shades of color" can be colors within the same wavelength interval but with different wavelengths.

[0046] Images of seed regions / parts can be used as input to a trained neural network or for training the network. It should be understood that an initial image with multiple seeds, rather than a single seed image, can alternatively be used as input to or to train a neural network. However, by extracting a single seed for input, the amount of training data for the neural network increases, thus increasing the likelihood of training the neural network to higher accuracy.

[0047] To train the neural network, the training images described above (including images of a single seed in this example) are input into the network. Coverage values ​​attributable to each image (e.g., values ​​determined through visual inspection) can also be used in the neural network, for example, as labels for the images. The neural network can process the images to estimate the coverage values ​​associated with each image. For example, given a training dataset such as the training images, forward propagation can sequentially compute the outputs in each layer and propagate the function signal forward through the network. In the final output layer, an objective loss function can measure the error between the inferred output and the given label (e.g., the coverage values ​​attributable to the training images). To minimize the training error, backpropagation can use the chain rule to backpropagate the error signal and compute the gradient with respect to all weights throughout the neural network. The weight parameters can then be updated using an optimization algorithm based on stochastic gradient descent (SGD).

[0048] A neural network can be a convolutional neural network (CNN). A neural network can be a deep neural network (DNN). A neural network can include multiple layers (an array of layers). For example, a neural network can include an input layer, 2D convolutional layers, batch normalization layers, ReLU layers, max pooling layers, fully connected layers, and an output layer, such as a regression layer. A neural network can include any of these layers, with any number and any suitable order between the input and output layers. The input to a neural network can be the image described above, and the output can be a value related to the coating coverage on the seed, such as the abrasion resistance of the coating, the percentage of coverage, or the uniformity of the coating thickness. A neural network can include a set of options for training using stochastic gradient descent with momentum. Once training is complete, the model can be re-optimized with a new set of options. For example, each time the model weights are updated in an optimization step, the initial learning rate can be set to change less in response to the estimation error.

[0049] The following section outlines an example of the neural network used in this method.

[0050] Neural networks can include an input layer for feeding 2D images into the network and applying data normalization. An image input layer can be created based on the resolution of a color image (e.g., an RGB image). For example, for a 100×100 resolution color image, where each image has 3 RGB values, the input layer could have 100 input layers and 3 (30,000) neurons.

[0051] The neural network may also include a first 2D convolutional layer that applies a sliding convolutional filter to the input. This layer convolves the input by moving the filter vertically and horizontally along the input, calculating the dot product of the weights and the input, and then adding a bias term.

[0052] Neural networks may also include a first batch normalization layer, which normalizes each input channel on a mini-batch. Using batch normalization layers between layers can accelerate the training of convolutional neural networks and reduce sensitivity to network initialization. The batch normalization layer first normalizes the activation of each channel by subtracting the mini-batch mean and dividing by the mini-batch standard deviation. Then, the layer shifts the input by a learnable offset β and scales it by a learnable scaling factor γ.

[0053] The neural network may also include a first ReLU layer, which performs a thresholding operation on each element of the input, where any value less than zero is set to zero.

[0054] Neural networks can also include max pooling layers, which perform downsampling by dividing the input into rectangular pooling regions and computing the maximum value of each region. For example, a max pooling layer can have a pool size [2 2] and a stride [2 2] (where the stride is the stride size of the vertical and horizontal traversal of the input, specified as a vector of two positive integers [ab], where a is the vertical stride size and b is the horizontal stride size).

[0055] The neural network can further include a second 2D convolutional layer, which has a similar structure to the first layer, except that the number of layers can be doubled.

[0056] The neural network may further include a second batch normalization layer, a second ReLU layer, a second max pooling layer, a third batch normalization layer, and a third ReLU layer.

[0057] Neural networks can also include a first fully connected layer that multiplies the input by a weight matrix and then adds a bias vector. The output size can be specified for the fully connected layer. For example, the output size can be the same as the size of a second or third 2D convolutional layer.

[0058] The neural network may include a fourth ReLU layer and a second fully connected layer. The size of the second fully connected layer may be 1 (e.g., it may have a value output from the second fully connected layer). The output of the second fully connected layer may be a value related to the seed coating coverage (e.g., abrasion resistance) shown in the input image.

[0059] Neural networks can also include regression layers that compute the semi-mean squared error loss for regression problems.

[0060] A set of options for training a neural network using stochastic gradient descent with momentum can be determined. For example, the number of epochs (corresponding to the number of epochs in which the data is fully traversed) can be set to 250. The execution environment or hardware resources used to train the network can use local or remote parallel pools. For example, computation / modeling and software development can be performed using parallel computing on a computer with multiple GPUs and CPUs (e.g., 4 GPUs and 12 CPUs). The training data can also be divided into mini-batches, for example, at a certain time in each epoch, with a size of 100. The initial learning rate (how much the model changes in response to the estimation error each time the model weights are updated) can be set to a value of, for example, 0.001. Regularization can be used to improve the performance of the model. The regularization factor can be set to 0.004. The data can be shuffled before each training epoch and before each network validation. Validation data can be used to validate the neural network during training. The validation accuracy and validation loss of the validation data can be calculated during training. Validation data can be images with their corresponding coverage values ​​(validation images set separately from the training images) in addition to the training data. Before network training stops, the number of iterations of the validation set loss can be greater than or equal to the number of iterations of the previous minimum loss, which can be set to 7. A training progress graph can be plotted during training.

[0061] Validation data (data not used to train the neural network, such as images) can be used to evaluate how well the neural network assesses seed coverage. For evaluation, validation data is fed into a trained neural network, and the network outputs a value for seed coverage or abrasion resistance. Evaluation can then be performed by comparing the abrasion resistance / coverage value determined by the trained neural network with a visually determined value for abrasion resistance / coverage. Validation of the neural network may include calculating the root-mean-square error (RMSE) of the visually determined coverage value and the coverage value predicted by the neural network. The trained neural network with the lowest RMSE is the one with the optimal options and architecture. Various configurations of the neural network can be implemented, and the trained neural network with the lowest RMSE can be selected as the trained neural network used in this method.

[0062] Once the neural network has been trained, the model can be re-optimized. Bayesian optimization can be performed by minimizing the classification error on the validation set, allowing for the selection and tuning of the neural network's architecture and options. A different set of options can be implemented. For example, this set of options may typically use the same values ​​described in the examples above, or different values ​​may be used. For instance, the initial learning rate could be reduced by a factor of 10 to 0.0001. By reducing the initial learning rate during re-optimization, the model can be changed less in response to the estimation error each time the model weights are updated, thus improving the resulting value representing seed coverage. Model validation can be performed using validation data (the same or different validation data as described above) to evaluate the model's performance, for example, to determine the model's accuracy.

[0063] The neural networks outlined above include structures and options that the inventors identified as particularly effective in calculating coating coverage on seeds.

[0064] In one example, over 500 images of coated, abraded, and unprocessed corn seeds (with varying qualities, resolutions, and light intensities) were collected and used as training images (Table 1) to train the neural network. Additionally, approximately 10% of the images (46 images) were used as validation images (27 images for red-coated seeds and 16 images for blue-coated seeds). Each of these images underwent visual inspection to assign a value to each image representing the amount of coating on the seed or the abrasion resistance of the coating.

[0065] Table 1

[0066] Images of coated seeds Images of worn seeds Images of raw seeds Corn (red husk) 112 224 22 Corn (blue coating) 82 164 22

[0067] Therefore, approximately 50 images were selected to validate the trained neural network. Figure 4 An example verification image of the corn (yellow in color) to be evaluated is shown. Figure 4 a) shows seeds coated with red. The seeds in this image have been exposed to mechanical stress, hence the presence of red areas in varying shades and yellow areas in varying shades. Figure 4 b) shows seeds coated with red, which have not been exposed to mechanical stress and therefore typically include a red coating of the same hue. Figure 4c) shows uncoated seeds (unprocessed seeds), and therefore are yellow. 4d) shows seeds coated with a blue coating. The seeds in this image have been exposed to mechanical stress, and therefore include blue areas of varying shades and yellow areas of varying shades. Each individual seed was assigned a value related to the coating coverage on the seed using a neural network; the average value represents the final score for abrasion resistance. Table 2 below lists the abrasion resistance for each image 4(a) through 4(d), determined both by visual inspection of the seeds and by a trained neural network. These values ​​are within acceptable ranges.

[0068] Table 2

[0069]

[0070] Figure 5 Examples of 25 seed portions extracted from images are shown, with visually determined abrasion resistance (labeled "expected") and abrasion resistance predicted by a trained neural network (labeled "predicted"). Figure 5 Seed regions with different colored coatings, coverage (thickness and percentage), lighting, etc., are shown. In particular, these images are of yellow corn with blue or red coatings. Figure 6 Images used for verification are shown (including) Figure 5 The graph shows the predicted abrasion resistance relative to the measured abrasion resistance (as shown in the image). In this example, the correlation between the measured and predicted abrasion resistance is 0.9455. More specifically, the correlation is 0.91 for the red coating and 0.98 for the blue coating. Therefore, it is clear that the trained neural network can provide a sufficiently accurate assessment of abrasion resistance.

[0071] Figure 7-10 They are respectively Figure 2-5 The color version.

[0072] In any of the above aspects, the various features can be implemented in hardware or as software modules running on one or more processors. A feature of one aspect can be applied to any other aspect.

[0073] The present invention also provides a computer program or computer program product for performing any of the methods described herein, and a computer-readable medium having a program thereon for performing any of the methods described herein. The computer program embodying the invention may be stored on a computer-readable medium, or it may be in the form of, for example, a signal, such as a downloadable data signal provided from an Internet website, or it may be in any other form.

[0074] Computing devices (such as data storage servers) can implement this invention and can be used to implement the methods of embodiments of this invention. The computing device may include a processor and memory. The computing device may also include a network interface for communicating with other computing devices, such as those according to embodiments of this invention.

[0075] For example, an implementation may consist of a network of such computing devices. The computing devices may also include one or more input mechanisms, such as a keyboard and mouse, and a display unit, such as one or more monitors. These components can be interconnected via a bus.

[0076] Memory may include computer-readable media, which can refer to a single medium or multiple media (e.g., a centralized or distributed database and / or associated caches and servers) configured to carry computer-executable instructions or to store data structures thereon. Computer-executable instructions may include, for example, instructions and data that can be accessed and caused to perform one or more functions or operations by a general-purpose computer, a special-purpose computer, or a special-purpose processing device (e.g., one or more processors). Therefore, the term "computer-readable storage medium" may also include any medium capable of storing, encoding, or carrying a set of machine-executable instructions and causing the machine to perform any one or more methods of the present invention. Thus, the term "computer-readable storage medium" may include, but is not limited to, solid-state memory, optical media, and magnetic media. By way of example and not limitation, such computer-readable media may include non-transitory computer-readable storage media, including random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage or other magnetic storage devices, and flash memory devices (e.g., solid-state storage devices).

[0077] A processor can be configured to control a computing device and perform processing operations, such as executing code stored in memory to implement the various methods described in the claims herein. The memory can store data read and written by the processor. As referred to herein, a processor can include one or more general-purpose processing devices, such as a microprocessor, a central processing unit, etc. A processor can include a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets or combinations of instruction sets. A processor can also include one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), a network processor, etc. In one or more embodiments, the processor is configured to execute instructions that implement the operations and steps discussed herein.

[0078] The display unit can show a representation of the data stored on the computing device, and can also display a cursor, dialog boxes, and screens, enabling users to interact with the programs and data stored on the computing device. The input mechanism allows users to input data and commands into the computing device.

[0079] A network interface (Network I / F) can connect to a network such as the Internet and can connect to other computing devices via the network. The Network I / F 997 can control data input / output to / from other devices via the network. Peripherals such as microphones, speakers, printers, power supplies, fans, chassis, scanners, trackball mice, etc., can be included in the computing device.

[0080] Embodiments of the invention, along with their various features and advantageous details, are explained more fully with reference to the non-limiting examples described and / or illustrated in the accompanying drawings and detailed in the following description. It should be noted that the features shown in the drawings are not necessarily drawn to scale, and those skilled in the art will recognize that features of one embodiment can be used with other embodiments, even if not explicitly stated herein. Descriptions of well-known components and processing techniques may be omitted to avoid unnecessarily obscuring the exemplary aspects of the invention. The examples used herein are merely for the purpose of facilitating understanding of how the invention can be implemented and further enabling those skilled in the art to practice the invention. Therefore, the examples herein should not be construed as limiting the scope of embodiments of the invention, which is defined only by the appended claims and applicable law.

[0081] It should be understood that embodiments of the present invention are not limited to the specific methods, protocols, devices, apparatuses, materials, applications, etc., described herein, as these can vary. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the claimed embodiments. It must be noted that, as used herein and in the appended claims, the singular forms “a,” “an,” and “described” include plural references unless the context clearly indicates otherwise.

[0082] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which embodiments of the invention pertain. Preferred methods, apparatus, and materials are described, although any methods and materials similar to or equivalent to those described herein may be used in practice or testing of embodiments.

[0083] Although only a few exemplary embodiments have been described in detail above, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the embodiments of the invention. The above-described embodiments of the invention can be advantageously used independently of any other embodiment, or in any feasible combination with one or more other embodiments.

[0084] Therefore, all such modifications are intended to be included within the scope of embodiments of the invention as defined in the following claims. In the claims, the means plus function clause is intended to cover not only structural equivalents but also equivalent structures as described herein for performing that function.

[0085] Furthermore, any reference numeral enclosed in parentheses in one or more claims should not be construed as limiting the claims. The words “comprising” and “including” do not exclude the presence of any element or step other than those listed in the claims or specification. A singular reference to an element does not exclude a plural reference to that element, and vice versa. One or more embodiments can be implemented by hardware comprising several different elements. In a device or apparatus claim that enumerates several means, several of these means can be implemented by the same hardware. The fact that certain measures are referenced in mutually different dependent claims does not mean that a combination of these measures cannot be used advantageously.

Claims

1. A method for processing a seed image, the method comprising: A seed image, including at least a portion of the seed, is fed into a trained neural network to generate a value related to the coating coverage on the seed. The trained neural network has been trained to generate values ​​related to the coating coverage on a seed using multiple training images, each training image including at least a portion of the training seed, and wherein each training image is labeled with a value related to the coating coverage on the training seed. The values ​​associated with the coating coverage on the seeds and the values ​​associated with the coating coverage on the training seeds represent the abrasion resistance of the coating.

2. The method of claim 1, wherein the value associated with the coating coverage on the seeds and / or training seeds represents at least one of the following: the percentage of the seed surface covered by the coating; the uniformity of coating coverage on the seeds.

3. The method as described in claim 1, wherein the method further includes the step of training the neural network.

4. The method of claim 1, wherein the trained neural network is trained using a first training image containing a training seed with a first color coating and a second training image containing a training seed with a second color coating.

5. The method of claim 4, wherein inputting the seed image includes inputting a seed image containing a seed with a first color coating.

6. The method of claim 1, wherein the seed image is generated by extracting a region of the seed from an image comprising a plurality of seeds.

7. The method of claim 6, wherein an algorithm for detecting circles in an image is used to extract the region of the seed, to detect the region within the seed, and then the region within the seed is extracted.

8. The method of claim 1, wherein the training image is generated by extracting regions of training seeds from an image comprising a plurality of training seeds.

9. The method of claim 8, wherein, An algorithm for detecting circles in an image is used to extract the region of the training seed, thereby detecting the region within the training seed and then extracting the region within the training seed.

10. The method of claim 1, wherein multiple seed images generated from the same image or taken from the same sample are input into a trained neural network to generate multiple values ​​related to the coating coverage of the corresponding seed images, and the average of the multiple values ​​is taken.

11. The method as described in claim 1, wherein, The neural network is at least one of the following: a convolutional neural network; and a deep neural network.

12. The method of claim 1, wherein the neural network comprises at least one of the following: an input layer, a 2D convolutional layer, a batch normalization layer, a ReLU layer, a max pooling layer, a fully connected layer, and a regression layer.

13. A computer program, when executed by a computing system including processor hardware and memory hardware, causes the processor hardware to perform the method as described in any one of claims 1 to 12.

14. An apparatus comprising processor hardware and memory hardware, the memory hardware storing processing instructions that, when the processor hardware executes the processing instructions, cause the processor hardware to perform the method as described in any one of claims 1 to 12.

Images